Editly

Run Z-Image Online, No GPU Required

Z-Image is the 6-billion-parameter, Apache 2.0 image model from Alibaba's Tongyi MAI team. Editly hosts it for text-to-image, so you can skip the 16GB VRAM checklist and the ComfyUI graph: write a prompt, choose one of five aspect ratios, and spend 3 credits per render.

Start creating

Running Z-Image Locally or Online: Pick Your Trade-Off

Almost everyone searching for Tongyi Z-Image lands on a HuggingFace checkpoint next to a ComfyUI tutorial. Both routes are legitimate. They simply charge you in different currencies, so here is the honest split before you clear 20GB of disk space.

Self-hosted
On EditlyEditly
What you needA CUDA card, Python, and a recent diffusers or ComfyUI buildA browser and an account
VRAMTurbo fits in 16GB at bfloat16; stable-diffusion.cpp documents 4GBNone of yours; inference runs server-side
Setup timeWeights download plus environment debuggingSign in and type
Sampler controlSteps, guidance, scheduler, negative prompt, seedPrompt and aspect ratio; sampling preset is fixed
LoRAs and fine-tuningThe base checkpoint is fine-tunable; Turbo is marked not fine-tunableNot available
Image editingWaiting on the Edit checkpoint releaseText-to-image only
Cost shapeHardware and electricity up front, renders after that3 credits per image, no hardware
Weights licenceApache 2.0, yours to keepApache 2.0, same model, hosted

Why Z-Image Punches Above Its 6B Weight Class

Bilingual Type Rendered Inside the Frame

Z-Image 6B was built to draw English and Chinese text directly into the picture, and that is the capability Tongyi MAI leads with in its own showcase. Poster headlines, shop signage, packaging copy and mock UI labels come out of the model instead of being pasted on in a design tool afterwards.

Eight Function Evaluations, Courtesy of Decoupled-DMD

Z-Image Turbo is a distilled checkpoint that reaches a finished picture in 8 function evaluations with classifier-free guidance switched off entirely. The distillation method behind it, Decoupled-DMD, is published as its own paper (arXiv:2511.22677), so the step count is documented engineering rather than a launch-day slogan.

Apache 2.0 Weights, Not an API-Only Model

Both checkpoints sit under the Apache 2.0 licence on HuggingFace and ModelScope. Download them, fine-tune the base model, redistribute your derivative, use the output commercially. That is a materially different arrangement from closed image models you can only reach through somebody else's endpoint.

What the Z-Image Form Actually Asks You For

  1. 01

    Open the image generator

    Go to the AI image tool and select Z-Image from the model list. Nothing to download, no Python environment, no building diffusers from source.

  2. 02

    Describe the shot

    Prompts run up to 1,000 characters. The model rewards concrete description: name the subject, the lighting, the camera framing, and put any text you want rendered in quotation marks.

  3. 03

    Pick a ratio and render

    Five aspect ratios are available here: 1:1, 4:3, 3:4, 16:9 and 9:16. Each render costs 3 credits and drops straight into your library.

Z-Image Turbo vs Base vs Edit: One 6B Family

Tongyi MAI ships one 6B family with four members, and search results mix them up constantly. A fourth checkpoint, Z-Image Omni-Base, is announced as a raw generation-and-editing base for community fine-tuning. The three below are the ones people actually search for, read straight from the official model zoo.

Z-Image Turbo
Base
Edit
StatusReleased 26 Nov 2025Released 27 Jan 2026Announced, not yet released
Sampling steps850 (28-50 recommended)50
Classifier-free guidanceOff, guidance 0On, guidance 3.0-5.0On
TaskText-to-imageText-to-imageInstruction-based editing
Visual quality / diversityVery high / lowHigh / mediumHigh / medium
Fine-tunableNoYesYes
Negative promptsNot usedStrongly recommendedNot documented yet

Z-Image: Licence, Hardware and the Turbo Split

Try Z-Image without installing anything

Three credits, five aspect ratios, and the same open-weight 6B model the ComfyUI crowd is busy downloading.

Generate with Z-Image